Pyrosphere

Pyrosphere helps fire brigades respond to fires quicker and have faster resource deployment with built-in fire reporting, resource management, and artificial intelligence monitoring.

Comment

Inspiration

California is a literal hotspot: every month, fires burn down farmland, forests, residential neighborhoods, and properties. While there are fire-detection systems in place, the advent of artificial intelligence allows for 24/7 monitoring of fire-risk zones and predictions (intensity, trajectory, etc.) based on current weather conditions.

What it does

  • Utilizes existing infrastructure of 1150 cameras across California and processes them through an ML model built on a dataset of more than 21,000 images to instantly identify smoke and fires.
  • Uses a fire prediction model that takes in weather conditions (wind, pressure, humidity, temperature) to predict fires: trained on 14,000 data entries.
  • Has a built-in reporting tool that allows for easier identification of people in need.
  • Makes resource management easier by displaying fire brigades, equipment and stations in the area.
  • Predicts the likelihood of future fires before they occur.

How we built it

  • NodeJs & FastAPI: Utilizes two backends suited for each task.
  • React, Typescript, HTML: Creates an interactive, sleek, and efficient interface.
  • YoloV8 Model: Computer vision model trained on the D-Fire Dataset with over 21,000 images.
  • MongoDB: Used for saving reports from users.

Challenges we ran into

  • Hardware limitations: Training such a large dataset leads to memory issues! So does running a big ML model on laptops.
  • Understanding UI for first-responders: The UI needs to be simple and effective.
  • Webscraping cameras: Finding access to 1150 cameras required a deep-dive into Chrome DevTools.
  • The time crunch: Working under a time constraint means we missed a lot of UI and backend features!

Accomplishments that we're proud of

  • Clean UI: It's our best work yet! Pyrosphere has a simple, yet effective design.
  • A really good ML model: As shown in live-video demos, it works really well.
  • The map: The ML processed camera previews look really cool!

What we learned

  • Machine learning: We gained a lot of experience training such a big model!
  • UI Design: We learned how to approach UI with a different audience.
  • Fires are unpredictable: Even with ML, we had our challenges in predicting fire behavior.
  • More can be done: There are so many more resources that can be applied to attack fires.
  • Prevention: There is a lot of work towards detecting fires, but more can be done to prevent them!

What's next for Pyrosphere

  • Finishing the UI: We were time-limited, so unfortunately some pages were left blank!
  • Trajectory prediction: We plan on using another model to predict fire trajectory.
  • Fire-impact zoning: We want to identify areas where fires are going to be most harmful based on previous data.

Check out the Google Drive links to see how well our model performs on frame-by-frame footage!

Share this project:

Updates